Papers of the day   All papers

Meta-Learning Symmetries by Reparameterization

Comments

Chelsea Finn: Convolution is an example of structure we build into neural nets. Can we _discover_ convolutions & other symmetries from data? Excited to introduce: Meta-Learning Symmetries by Reparameterization https://arxiv.org/abs/2007.02933 w/ @allan_zhou1 @TensorProduct @StanfordAILab Thread👇

6 replies, 907 likes


Allan Zhou: A method for meta-learning the "structure" of equivariant operations, separately from the filter parameters. This helps preserve equivariances meta-learned from data augmentation when solving new tasks, without needing any new augmented data!

0 replies, 2 likes


Arman: First they came after our model selection with auto-ml, then hyperparam tuning with xgboost that doesn't even need it, then feature engineeing with DL, and now we don't even need to design a CNN, data will tell you when it needs one!? ML community needs a union to protest! ;-)

0 replies, 1 likes


Content

Found on Jul 08 2020 at https://arxiv.org/pdf/2007.02933.pdf

PDF content of a computer science paper: Meta-Learning Symmetries by Reparameterization